Figure 4 illustrates the proposed architecture of the conceptual system, which consists of four information processing layers and three vertical subsystems, namely, perception, central processing, and action according to Sloman’s H-Cogaff scheme. The lowest horizontal layer above the distributed control system (DCS) contains semi-autonomous agents that represent different levels of data abstraction and information processing mechanisms of the system.
The middle two layers (i.e., the reactive and deliberative layers) interact with the external 479 Multi-agent Systems for Industrial Applications: Design, Development, and Challenges
Fig. 3. Human cognition and affect (H-Cogaff) architecture Sloman & Scheutz (2002)
environment via the DCS and thus the industrial process by acquiring perceptual inputs and generating actions. The perceptual and action subsystems are divided into several layers of abstraction to function effectively. This can be achieved, for example, by categorizing observed events at several levels of abstraction, and allowing planning agents to generate behavior (actions) in a hierarchically organized manner.
IC A M system
M eta m anager
D eliberative BB m anager
N orm al event Supervisor
A bnorm al event supervisor
R eactive BB m anager
Plan executer A ction BB m anager
Prim ary plan scheduler
Fault detection &
isolation # 2 Process optim izer
M odel identification Perception BB
m anager
Secondary plan scheduler Fault detection &
isolation # 1 W ireless sensor
netw ork m anager
D ata statistical pre-processor
D ata reconciliation D base m anager
U ser interface m anager
M anufacturing environm ent A ctuators W ireless
Sensors
D ata flow C om m unication & control flow BB: Blackboard
Fig. 4. ICAM system architecture and behavior model
480 Multi-Agent Systems - Modeling, Control, Programming, Simulations and Applications
The basic flow of control in the system begins when perceptual input arrives at the lowest level in the architecture. If the reactive layer can deal with this input then it will do so, otherwise, bottom-up activation will occur and control will be passed to the deliberative layer. If the deliberative layer can handle the situation then it will do so, typically by making use of top-down execution of reactive agents. Otherwise, it will pass control to the meta-management layer to resolve any internal conflicts in the architecture or notify the operator that it cannot do so. As illustrated in figure 4, the functionalities of the ICAM system agents in each layer are.
• Statistical pre-processing agent to clean the measured data from undesired discrepancies such as missed data and outliers.
• Data reconciliation agent to reconcile measured data according to mass and energy conservation laws.
• Database agent to store real-time data for historical and other purposes.
• Several fault detection and isolation (FDI) agents to detect any abnormal events using different FDI approaches.
• An optimization agent to generate the optimal operating plans.
• a system identification agent to identify the mathematical model of the industrial process.
• Two intelligent supervisory agents to manage the industrial plant during normal and abnormal situations.
• A meta manager agent to manage and coordinate the different system agents.
• A set of black board agents to facilitate asynchronous communications among the system agents.
• A set of task scheduling agents to execute the required plan according to different time frames.
Rigorous coordination of the behavior of the ICAM system layers and agents is crucial to success. A sound coordination scheme will allow us to assess its performance, and to evaluate how the internal agents of the system interact when certain internal/external events occur. Furthermore, it permits system behavior modeling to simulate the most critical design characteristics such as concurrency, autonomy, task distribution and parallelism, in order to guarantee robust and coherent performance. Due the complexity of modern manufacturing plants, intelligent systems (e.g., ICAM) have to be distributed, which makes the coordination of such systems very difficult and challenging.
Durfee et al. (Durfee & Montgomery, 1991) proposed an informal theory that integrates organizational behavior, long term plans, and short term schedules into one coordination framework, and treats coordination as a distributed search process through the hierarchical space of the possible interacting behaviors of the individual agents to find a collection that satisfactorily achieves the agents’ goals. The theory emphasizes several topics such as:
• hierarchical behavior representation to express different dimensions of behavior at different levels of detail,
• metrics for measuring the quality of coordination between agents,
• distributed search protocol for guiding the exchange of information between agents during the distributed search,
481 Multi-agent Systems for Industrial Applications: Design, Development, and Challenges
• local search algorithm for generating alternative behaviors at arbitrary levels of abstractions, and
• control knowledge and heuristics for guiding the overall search process to improve coordination.
Durfee also suggested that introducing a meta-level organization in the intelligent system to manage coordination between agents, and separating knowledge representation into domain-level and meta-level types, would enhance coordination and make it more robust.
Agents use domain-level knowledge to influence what goals they pursue, and use meta-level knowledge to decide how, when, and where to form and exchange behavioral models (Durfee et al., 1989). Durfee’s informal theory and suggestions give the big picture of how agents should coordinate their activities within an intelligent system or even a society of intelligent agents. So far we have addressed the knowledge and organization separation issues by adopting the H-CogAff architecture proposed by Sloman. ICAM interacts with the external world through its reactive and deliberative agents, whereas the meta-level layer dictates the internal behavior of the system. Furthermore, domain-level knowledge is encoded in the deliberative agents and the meta-level knowledge is encoded in the self reflective layer.
As illustrated by figure 4, the proposed conceptual behavior model of the ICAM system was built upon our previous work in which we defined the architecture of the system, its functional modules, and its coordination mechanisms (Taylor & Sayda, 2005a;b). We adopted Sloman’s H-Cogaff architectural scheme because it met most of our system requirements (Sloman, 2001).
The behavioral model was drawn as a page hierarchy to make it compatible with hierarchical colored petri net (HCPN) terminology, which could be used to analyze the logical correctness and the dynamic behavior of the system; however, this has not been done.
The prime page in the model is called ICAM which contains all the subpages of the system.
Each subpage represents an independent agent which interacts with others by means of communications (represented by thin bidirectional arrows). Other agents may process data received from the plant directly (data flow is represented by open thick unidirectional arrows).
The meta manager is the main coordinator of the whole system, which guarantees more robust and coherent performance. The meta manager is basically a rule-based expert system, which codifies all possible system behaviors and agent interactions as a behavior hierarchy in its knowledge base. Agent behavior is represented in the behavior hierarchy by a single structure, which will use the same message structure communicated between agents. This will result in a better system performance. Table 1 illustrates the unified behavior conceptual structure.
Field name Field content
Tag Message ID
From Sender
To Recipient
What Goals
How Plans
When Schedule
How long Task length
Why Meta reasoning
Table 1. Conceptual structure of behavioral message
482 Multi-Agent Systems - Modeling, Control, Programming, Simulations and Applications